Accurate and Reliable Prediction of Relative ... - Columbia University


Accurate and Reliable Prediction of Relative...

3 downloads 87 Views 3MB Size

Article pubs.acs.org/JACS

Accurate and Reliable Prediction of Relative Ligand Binding Potency in Prospective Drug Discovery by Way of a Modern Free-Energy Calculation Protocol and Force Field Lingle Wang,† Yujie Wu,† Yuqing Deng,† Byungchan Kim,† Levi Pierce,† Goran Krilov,† Dmitry Lupyan,† Shaughnessy Robinson,† Markus K. Dahlgren,† Jeremy Greenwood,† Donna L. Romero,‡ Craig Masse,‡ Jennifer L. Knight,† Thomas Steinbrecher,† Thijs Beuming,† Wolfgang Damm,† Ed Harder,† Woody Sherman,† Mark Brewer,† Ron Wester,‡ Mark Murcko,† Leah Frye,† Ramy Farid,† Teng Lin,† David L. Mobley,⊥ William L. Jorgensen,∥ Bruce J. Berne,§ Richard A. Friesner,§ and Robert Abel*,† †

Schrödinger, Inc., 120 West 45th Street, New York, New York 10036, United States Nimbus Discovery, 25 First Street, Suite 404, Cambridge, Massachusetts 02141, United States § Department of Chemistry, Columbia University, 3000 Broadway, New York, New York 10027, United States ∥ Department of Chemistry, Yale University, New Haven, Connecticut 06520, United States ⊥ Departments of Pharmaceutical Sciences and Chemistry, University of CaliforniaIrvine, Irvine, California 92697, United States ‡

S Supporting Information *

ABSTRACT: Designing tight-binding ligands is a primary objective of small-molecule drug discovery. Over the past few decades, freeenergy calculations have benefited from improved force fields and sampling algorithms, as well as the advent of low-cost parallel computing. However, it has proven to be challenging to reliably achieve the level of accuracy that would be needed to guide lead optimization (∼5× in binding affinity) for a wide range of ligands and protein targets. Not surprisingly, widespread commercial application of free-energy simulations has been limited due to the lack of large-scale validation coupled with the technical challenges traditionally associated with running these types of calculations. Here, we report an approach that achieves an unprecedented level of accuracy across a broad range of target classes and ligands, with retrospective results encompassing 200 ligands and a wide variety of chemical perturbations, many of which involve significant changes in ligand chemical structures. In addition, we have applied the method in prospective drug discovery projects and found a significant improvement in the quality of the compounds synthesized that have been predicted to be potent. Compounds predicted to be potent by this approach have a substantial reduction in false positives relative to compounds synthesized on the basis of other computational or medicinal chemistry approaches. Furthermore, the results are consistent with those obtained from our retrospective studies, demonstrating the robustness and broad range of applicability of this approach, which can be used to drive decisions in lead optimization.



affinities).1,2 The most rigorous approach to this problem is free-energy simulation. A variety of free-energy simulation methods, such as free-energy perturbation (FEP), thermodynamic integration (TI), and λ dynamics, employ an analysis of atomistic molecular dynamics or Monte Carlo simulations to determine the free-energy difference between two related ligands via either a chemical or alchemical path.3−9 In drug discovery lead optimization applications, the calculation of relative binding affinities (i.e., the relative difference in binding energy between two compounds) is generally the quantity of interest and is thought to afford significant reduction in

INTRODUCTION

Protein−ligand binding is central to both biological function and pharmaceutical activity. Some ligands simply inhibit protein function, while others induce protein conformational changes and hence can modulate key cell-signaling pathways. In either case, achieving a desired therapeutic effect is dependent upon the magnitude of the binding affinity of ligand to target receptor. Designing tight-binding ligands while maintaining the other ligand properties required for safety and biological efficacy is a primary objective of small-molecule drug discovery projects. A principal goal of computational chemistry and computeraided drug design (CADD) is therefore the accurate prediction of protein−ligand free energies of binding (i.e., binding © 2015 American Chemical Society

Received: December 15, 2014 Published: January 27, 2015 2695

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703

Article

Journal of the American Chemical Society

ology,30,31 where a substantial number of bond charge corrections for challenging chemistries have been developed. A comparison of the performance of OPLS2.1 relative to MMFF in reproducing quantum chemical torsional profiles and conformational energies is presented in Figure 1. Performance of the nonbonded interaction model has been

computational effort as compared to absolute binding freeenergy calculations. Nearly three decades have passed since the initial applications of free-energy methods to the calculation of protein−ligand binding affinities were first reported by the Jorgensen, McCammon, and Kollman groups.10−15 Subsequent efforts since that seminal work have reported anecdotal results for a small number of protein−ligand complexes, but these have suffered from a lack of computing power and inadequacies in both sampling algorithms and molecular mechanics force fields.1,5,8 As a result, use of free-energy calculations was limited in an industrial drug discovery setting, where high throughput, predictive accuracy, and robustness are required to make a significant impact. In recent years, FEP calculations have benefitted from improved force fields, new sampling algorithms, and the emergence of low-cost parallel computing, which have resulted in the level of accuracy and turnaround time needed to impact lead optimization efforts, as demonstrated in several academic projects.1,5,8,16−20 However, it has not been demonstrated that highly accurate results can be achieved reliably across a wide range of ligands and protein targets, as would be needed for the method to be useful in industrial pharmaceutical research programs. Here, we report an FEP protocol that enables highly accurate affinity predictions across a broad range of ligands and target classes (over 200 ligands and 10 targets). The ligand perturbations include a wide range of chemical modifications that are typically seen in medicinal chemistry efforts, with modifications of up to 10 heavy atoms routinely included. Critically, we have applied the method in eight prospective discovery projects to date, with the results from two of those projects disclosed in this work. The high level of accuracy obtained in the prospective studies demonstrates the ability of this approach to drive decisions in lead optimization.



Figure 1. Histograms of root-mean-square error in force field relative energies, evaluated (A) over one-dimensional torsional angle scans and (B) between conformational minima, on a set of 8365 compounds. Errors are established with respect to quantum mechanical LMP2/ccpVTZ(-f) energies evaluated on B3LYP/6-31G* optimized structures. The compound set is generated from a 6 million compound repository of druglike compounds where selected molecules are subsequently fragmented about rotatable torsion bonds retaining key proximate substituents. Selected fragment molecules are chosen such that their constituent rotatable bonds retain sufficient similarity to the OPLS2.1 training set. A rotatable bond is deemed sufficiently similar if the set of atom typed quartets across the bond match a member of the OPLS2.1 training set.

FREE-ENERGY PERTURBATION TECHNOLOGY AND METHODOLOGY

The achievement of the results mentioned above is the consequence of an improved force field (OPLS2.1), enhanced sampling, and an automated workflow to ensure that all results are reproducible and realizable with minimal user interaction. Over the past decades, force fields for proteins, nucleic acids, lipids, and other biological molecules have improved substantially via fitting parameters to quantum chemical and experimental data;21−25 however, adequate parametrical coverage for druglike molecules has lagged behind. For example, MMFF,26 a widely used force field, is trained against just 140 fragment-sized compounds representing typical organic moieties found in druglike molecules. Our analysis of 1 million purchasable druglike compounds indicates that on the order of tens of thousands of such compounds are required to represent the diversity of even this limited chemical space. Using the OPLS force field23−25 as a starting point, we have developed a new force field, OPLS2.1,27 that incorporates a robust model for nonbonded interactions (van der Waals parameters and partial charges) in conjunction with extensive training of torsional and covalent parameters against more than 10,000 representative organic compounds. In addition, missing parameters for any molecule can be generated via an automated algorithm that performs the appropriate quantum mechanics calculations and torsion fitting. The torsional parameters are obtained by constructing model compounds containing the relevant torsional structures and fitting the parameters to quantum chemical data computed at the LMP2/cc-pVTZ(-f) level of theory, which has been shown to yield accurate relative conformational energies for the systems being modeled.28,29 Ligand atomic partial charges are computed via CM1A-BCC method-

initially evaluated in the prediction of aqueous solvation free energies, the results of which were reported in a prior publication32 and are summarized in Table 1, along with a comparison to other widely used force fields. These calibration results suggest that OPLS2.1 provides robust force field coverage in the space of druglike ligands and represents a significant advance in this regard as compared to previous general organic ligand force fields. In addition to the development of accurate potential energy functions, a significant challenge in FEP calculations is ensuring that

Table 1. Error Statistics for Solvation Free-Energy Resultsa force field ChelpG/CharmM AM1-BCC/ GAFF OPLS 2005 OPLS2.1

MUE (kcal/mol)

RMSE (kcal/mol)

% > 2 kcal/mol error

1.93 1.17

2.28 1.39

44.7 15.0

1.10 0.73

1.33 0.88

8.5 2.1

a

Reported by Shivakumar et al.32 using OPLS2.1, OPLS2005, AM1BCC/GAFF, and ChelpG/CHARMM-MSI for the test set of 239 small molecules. MUE, mean unsigned error; RMSE, root-meansquare error.

2696

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703

Article

Journal of the American Chemical Society

Figure 2. (a) Whole FEP workflow for protein−ligand binding-affinity calculations. Key steps in the workflow, including mutation graph generation, REST region assignment, cycle closure convergence, and error estimates, are explicitly shown in the figure. Note that the FEP/REST enhanced sampling method and cycle closure error estimate were reported in prior publications,37,38 and important related work about perturbation graph generation was also reported in a prior publication;41 the remaining components of the workflow are novel. (b) Example of mapping of a perturbation space onto a set of pathways for thrombin ligands generated from the workflow. Each line represents two full FEP calculations, one conducted in the receptor and one in solution, each perturbing between two connected ligands. (c) Correlation plot of FEP-predicted and experimental binding affinities for thrombin ligands, as generated from the automated workflow. problematic torsional barriers, which can limit ergodicity, to be surmounted in a routine fashion. A crucial aspect of the use of FEP/REST in practical calculations is the selection of the REST region. Heuristic rules must be developed to determine which atoms in the ligand and the protein environment should be included in the enhanced sampling region. We have developed an automated algorithm to select the REST region, and this algorithm was employed in a uniform fashion in all the studies reported here. Details about the REST region selection algorithm are described in the Methods section in Supporting Information. Third, Desmond with FEP/REST has been implemented to run on graphics processing units (GPUs). The GPU implementation provides 50−100× speedup over a single central processing unit (CPU) and approximately 5−10× performance/cost improvement as compared to a commodity PC cluster.40 For a typical FEP calculation (∼6000 atoms in the protein) with the protocol described in this work, four perturbations per day can be completed by use of eight commodity Nvidia GTX-780 GPUs, making it feasible to evaluate thousands of molecules per year in the context of a drug discovery program with compute resources that are well within the reach of both academic institutions and commercial enterprises. We also note here, consistent with the experiences reported in ref 40, that GeForce cards do require a significant information technology (IT) commitment for effective use in a production setting. Another critical aspect of the FEP protocol described here is ease of use, which is essential in order to have a broad impact on drug discovery projects. When considering the modification of a lead molecule, one generally explores a space of possible perturbations at different positions, using a variety of substituents. Prior implementations of FEP methods have generally required large amounts of human time to set up the calculations, and this manual setup is error-prone. In

the molecular dynamics simulations provide a sufficiently complete sampling of the phase space of the system, while at the same time retaining computational tractability using currently available hardware. A wide range of approaches have been investigated in the literature with the objective of improving the ability to surmount energy barriers and/or reducing computation time in large-scale biomolecular simulations.16,33 The methodology we have developed retains a full treatment of all degrees of freedom in the system, using a number of methods to substantially enhance the efficiency of phase space exploration, and is readily automated and applicable to a significant fraction of practical drug discovery projects. The key elements of our solution to this problem are outlined below. First, we employ the Desmond program to run FEP simulations. Desmond provides good single-node and parallel performance, yielding superior performance/cost ratios as compared to alternatives.34 Second, we have augmented the molecular dynamics/replica exchange capabilities in Desmond with the newly developed FEP/ REST (free energy perturbation/replica exchange with solute tempering) algorithm.35−37 FEP/REST enables simulations of a selected subsystem with replicas in a higher effective temperature regime than the remainder of the system, and thus precisely focuses sampling efforts where needed to properly traverse the relevant phase space. Prior results of FEP/REST computations demonstrated a notable improvement in predicting relative ligand binding affinities for two ligands that bind to thrombin. This improvement was accomplished by effectively locally heating the binding region yet retaining rigorous Boltzmann sampling.37 Additional work demonstrated similarly improved binding predictions for CDK2 ligands38 and HIV-1 reverse transcriptase inhibitors.39 The REST methodology thus enables 2697

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703

Article

Journal of the American Chemical Society Table 2. Relative Binding Free-Energy Calculation Resultsa system no. of compds binding affinity range (kcal/mol) crystal structure series ref no. of perturbations MUE FEP RMSE FEP avg σcc obs R-value FEP P-value FEP obs R-value, MW obs R-value, MM-GB/SA obs R-value, Glide SP anticip FEP R-value anticip exptl R-value

BACE

CDK2

JNK1

MCL1

p38

PTP1B

thrombin

Tyk2

36 3.5 4DJW 46 58 0.84 ± 0.08 1.03 ± 0.08 0.65 0.78 ± 0.07 3.9 × 10−5 0.14 −0.40 0.00 0.64 ± 0.09 0.88 ± 0.03

16 4.2 1H1Q 47 25 0.91 ± 0.12 1.11 ± 0.12 0.57 0.48 ± 0.19 1.2 × 10−2 −0.48 −0.53 −0.56 0.73 ± 0.11 0.92 ± 0.03

21 3.4 2GMX 48 31 0.78 ± 0.12 1.00 ± 0.15 0.30 0.85 ± 0.07 7.0 × 10−8 −0.39 0.65 0.24 0.64 ± 0.12 0.88 ± 0.04

42 4.2 4HW3 49 71 1.16 ± 0.10 1.41 ± 0.12 0.91 0.77 ± 0.05 2.2 × 10−7 −0.55 0.42 0.59 0.71 ± 0.07 0.91 ± 0.02

34 3.8 3FLY 50 56 0.80 ± 0.08 1.03 ± 0.09 0.76 0.65 ± 0.09 1.6 × 10−7 −0.46 0.66 0.14 0.67 ± 0.08 0.89 ± 0.03

23 5.1 2QBS 51 49 0.89 ± 0.12 1.22 ± 0.17 0.94 0.80 ± 0.08 7.8 × 10−6 −0.84 0.67 0.55 0.79 ± 0.07 0.94 ± 0.02

11 1.7 2ZFF 45 16 0.76 ± 0.13 0.93 ± 0.15 0.93 0.71 ± 0.24 1.1 × 10−2 −0.48 0.93 0.53 0.37 ± 0.26 0.68 ± 0.15

16 4.3 4GIH 52,53 24 0.75 ± 0.11 0.93 ± 0.12 0.46 0.89 ± 0.07 2.3 × 10−7 0.00 0.79 0.79 0.74 ± 0.10 0.92 ± 0.03

a

Eight different receptors, covering a broad range of protein types, were studied. The number of ligands, experimental binding affinity range of ligands, crystal structure used in the simulation, original publication reporting the experimental binding affinity, and number of perturbations for each system are reported. Details about how the data set was selected, and how the experimental binding free energies were obtained, are included in Supporting Information. Several different metrics to assess the performance of FEP results including mean unsigned error (MUE) and root mean square error (RMSE) for all perturbations, correlation coefficient (R) between FEP-predicted binding affinities and experimental results, and average error for predictions calculated by cycle closure algorithm (avg σcc) are also reported. For comparison, MM-GB/SA and Glide SP scoring results are also reported. The FEP scoring weighted average R-value obtained is 0.75, for MM-GB/SA it is 0.35, and for Glide SP it is 0.29. Expected correlation coefficient between FEP-predicted binding affinities and experimental results (anticip FEP R-value) and expected correlation coefficient between two experimental measurements of binding affinities (anticip exptl R-value), with assumed RMSEs of 1.1 and 0.4 kcal/mol for FEP-predicted binding affinities and experimental data, respectively, are also shown (see details in Supporting Information). Errors for MUE, RMSE, and R values by use of the bootstrapping method are also reported. Free energies are in units of kilocalories per mole.

with the OPLS 2005 force field and a manual setup,38 and the remaining seven data sets were first studied here. Structures of the individual ligands and the target perturbations used as starting points for the FEP calculations in each data set, as well as other methodological details, are given in Supporting Information. A summary of the performance for all the pairs of perturbations is also provided in Table 2. The combination of high correlations with experimental binding affinity for each system and a low root-mean-square error (RMSE) for all 330 perturbations implies results of sufficient quality to drive decisions in the hit-to-lead and lead-optimization phases of drug discovery projects. Table 3 reports a binned error distribution for all 330 perturbations, indicating a roughly Gaussian distribution with a standard deviation of 1.1 kcal/mol.

the protocol described here, the user inputs the molecules of interest (in any supported standard format) into a graphical interface, and the perturbation pathways are automatically generated by a variant of the LOMAP mapping algorithm.41 In the LOMAP algorithm, the maximum common substructure (MCS) between any pair of compounds is generated and their similarity is measured. Then ligand pairs with high similarity scores are connected by edges, where each edge represents one FEP calculation that will be performed between the two ligands. The perturbation graph topology is also optimized such that (1) each edge will, if possible, be nested within at least one closed cycle; and (2) there will be at least one path containing fewer than five edges between any pair of compounds. Figure 2a shows the automated FEP workflow for protein−ligand binding free-energy calculations, and an example mapping of a ligand series onto a set of pathways is shown in Figure 2b. The 16 separate calculations shown in Figure 2b can be prepared in approximately 30 min, whereas manual setup without a graphical user interface and automated mapping protocols would take significantly longer. Finally, our approach includes an assessment of the reliability of the calculations, previously a notorious weak point of free-energy methods. The use of multiple pathways, via a cycle closure analysis, enables more reasonable sampling error estimates for the calculations.38 The estimated error provides an approximation of calculation precision, which is particularly important for the prospective use of the method. Note that force field errors cannot be addressed by any such approach; cycle closure analysis error estimates analyze sampling problems only, that is, they estimate the minimal error in the free-energy results based on the conformational space sampled in all simulations.

Table 3. Error Distribution for All 330 Perturbationsa



absolute error (kcal/mol)

anticip %

obs %

8 had an experimental pKi > 8. Key to labels: TP = true positive, FN = false negative, FP = false positive, TN = true negative.

These accurate affinity predictions in both projects allowed the teams to reliably deprioritize a large number of proposed compounds and to focus synthesis and assay resources on efficiently achieving project potency and ADME goals.

the project and over a period of several months, 195 compounds were prospectively scored with FEP, and 22 were synthesized and assayed. The results of these predictions are shown in Figure 5. In total, 156 of the 195 compounds were predicted to have pKi ≤ 8, and the other 39 were predicted to have pKi > 8. Fifteen of the compounds predicted to have pKi ≤ 8 were, despite the predictions, synthesized and assayed to test various ADME hypotheses. As shown in Figure 5, 14 of these compounds (93%) turned out to have pKi ≤ 8, as predicted (only one false negative was found). Given this true negative rate of 93%, it is expected that ∼145 of the 156 compounds predicted to have pKi ≤8 would have been true negatives, while only ∼11 (7%) would have been false negatives (pKi > 8). Of the seven synthesized compounds predicted to have pKi > 8, five did have an experimental pKi > 8 and two did not (71% true positive rate). This constitutes a 6-fold enrichment in the synthesis of tight-binding molecules: only 12% of the compounds that were not prioritized by the FEP calculations were found to have pKi > 8, while 71% (5 out of 7) of those compounds that were predicted by FEP were correctly predicted to fall in this range. Note that a significant fraction of the 119 compounds synthesized without the benefit of FEP scoring were expected to be potent on the basis of more conventional analyses. Thus, the observed 6-fold enrichment in the synthesis of tight-binding molecules provides suggestive evidence that FEP scoring provides a substantial reduction in false positives relative to compound synthesized on the basis of other approaches. The accuracy of the results in project II was similar to those observed in project I. In project I, the average error between predicted and experimental affinity for the 22 compounds was 1.1 kcal/mol (average pKi error of 0.8). In project II, the average error for 20 compounds that were prospectively predicted and experimentally assayed was 0.9 kcal/mol (average pKi error of 0.7). Thirty-seven compounds with pKi predictions ≤ 8 were not synthesized; the true negative rate in project II was 75% based on results for four compounds predicted to have pKi ≤ 8 that were subsequently synthesized.



CONCLUSIONS The work described here addresses several major challenges to using FEP in drug discovery programs. The OPLS2.1 force field has considerably greater chemical space coverage than other widely used force fields and provides sufficient energetic accuracy for meaningful prospective free-energy calculations. Furthermore, the efficient FEP implementation reported herein extends the Desmond/GPU molecular dynamics engine to incorporate REST enhanced sampling, thereby improving simulation convergence. The combination of improved force field and superior sampling method has contributed to improved accuracy of the FEP protocol. Additionally, the fully automated FEP calculation setup and simulation quality analysis reduce human error and workload, thus making the approach accessible to a broad population of researchers in drug discovery. For a typical-sized drug target, four perturbations can be completed by use of eight commodity Nvidia GTX-780 GPUs. The aggregate effect of these advances now positions free-energy calculations to play a guiding role in the hit-to-lead and lead-optimization phases, as indicated by encouraging results in the two active drug discovery projects presented here. The preceding notwithstanding, a highly accurate and robust FEP methodology is not, in any way, a replacement for a creative and technically strong medicinal chemistry team; it is necessary to generate the ideas for optimization of the lead compound that are synthetically tractable and have acceptable values for a wide range of druglike properties (e.g., solubility, membrane permeability, metabolism, etc.). Rather, the computational approach described here can be viewed as a tool to enable medicinal chemists to pursue modifications and new synthetic directions that would have been considered too risky without computational validation or to eliminate compounds that would be unlikely to meet the desired target 2701

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703

Article

Journal of the American Chemical Society

(10) McCammon, J. A.; Gelin, B. R.; Karplus, M. Nature 1977, 267, 585. (11) Jorgensen, W. L.; Ravimohan, C. J. Chem. Phys. 1985, 83, 3050. (12) Bash, P.; Singh, U.; Brown, F.; Langridge, R.; Kollman, P. Science 1987, 235, 574. (13) Kollman, P. Chem. Rev. 1993, 93, 2395. (14) Wong, C. F.; McCammon, J. A. J. Am. Chem. Soc. 1986, 108, 3830. (15) Merz, K. M.; Kollman, P. A. J. Am. Chem. Soc. 1989, 111, 5649. (16) Deng, Y.; Roux, B. J. Phys. Chem. B 2009, 113, 2234. (17) Gallicchio, E.; Lapelosa, M.; Levy, R. M. J. Chem. Theory Comput. 2010, 6, 2961. (18) Durrant, J.; McCammon, J. BMC Biol. 2011, 9, 1. (19) Riniker, S.; Christ, C.; Hansen, H.; Hü nenberger, P.; Oostenbrink, C.; Steiner, D.; van Gunsteren, W. J. Phys. Chem. B 2011, 115, 13570. (20) Michel, J.; Essex, J. W. J. Med. Chem. 2008, 51, 6654. (21) MacKerell, A. D., Jr.; Bashford, D.; Bellott, M.; Dunbrack, R. L., Jr.; Evanseck, J. D.; Field, M. J.; Fischer, S.; Gao, J.; Guo, H.; Ha, S.; Joseph-McCarthy, D.; Kuchnir, L.; Kuczera, K.; Lau, F. T. K.; Mattos, C.; Michnick, S.; Ngo, T.; Nguyen, D. T.; Prodhom, B.; Reiher, W. E., III; Roux, B.; Schlenkrich, M.; Smith, J. C.; Stote, R.; Straub, J.; Watanabe, M.; Wiórkiewicz-Kuczera, J.; Yin, D.; Karplus, M. J. Phys. Chem. B 1998, 102, 3586. (22) Wang, J.; Wolf, R. M.; Caldwell, J. W.; Kollman, P. A.; Case, D. A. J. Comput. Chem. 2004, 25, 1157. (23) Jorgensen, W. L.; Tirado-Rives, J. J. Am. Chem. Soc. 1988, 110, 1657. (24) Jorgensen, W. L.; Maxwell, D. S.; Tirado-Rives, J. J. Am. Chem. Soc. 1996, 118, 11225. (25) Kaminski, G. A.; Friesner, R. A.; Tirado-Rives, J.; Jorgensen, W. L. J. Phys. Chem. B 2001, 105, 6474. (26) Halgren, T. A.; Nachbar, R. B. J. Comput. Chem. 1996, 17, 587. (27) OPLS 2.1; Schrodinger, Inc.: New York, 2014. (28) Bochevarov, A. D.; Harder, E.; Hughes, T. F.; Greenwood, J. R.; Braden, D. A.; Philipp, D. M.; Rinaldo, D.; Halls, M. D.; Zhang, J.; Friesner, R. A. Int. J. Quantum Chem. 2013, 113, 2110. (29) Murphy, R. B.; Beachy, M. D.; Friesner, R. A.; Ringnalda, M. N. J. Chem. Phys. 1995, 103, 1481. (30) Storer, J.; Giesen, D.; Cramer, C.; Truhlar, D. J. Comput.-Aided Mol. Des. 1995, 9, 87. (31) Jakalian, A.; Jack, D. B.; Bayly, C. I. J. Comput. Chem. 2002, 23, 1623. (32) Shivakumar, D.; Harder, E.; Damm, W.; Friesner, R. A.; Sherman, W. J. Chem. Theory Comput. 2012, 8, 2553. (33) Jorgensen, W. L.; Tirado−Rives, J. J. Comput. Chem. 2005, 26, 1689. (34) Bowers, K. J.; Chow, E.; Xu, H.; Dror, R. O.; Eastwood, M. P.; Gregersen, B. A.; Klepeis, J. L.; Kolossvary, I.; Moraes, M. A.; Sacerdoti, F. D.; Salmon, J. K.; Shan, Y.; Shaw, D. E. In Proceedings of the 2006 ACM/IEEE conference on Supercomputing, Tampa, FL, 2006; p 84. (35) Liu, P.; Kim, B.; Friesner, R. A.; Berne, B. J. Proc. Natl. Acad. Sci. U.S.A. 2005, 102, 13749. (36) Wang, L.; Friesner, R. A.; Berne, B. J. J. Phys. Chem. B 2011, 115, 9431. (37) Wang, L.; Berne, B. J.; Friesner, R. A. Proc. Natl. Acad. Sci. U.S.A. 2012, 109, 1937. (38) Wang, L.; Deng, Y.; Knight, J. L.; Wu, Y.; Kim, B.; Sherman, W.; Shelley, J. C.; Lin, T.; Abel, R. J. Chem. Theory Comput. 2013, 9, 1282. (39) Cole, D. J.; Tirado-Rives, J.; Jorgensen, W. L. J. Chem. Theory Comput. 2014, 565. (40) Bergdorf, M.; Kim, E. T.; Rendleman, C. A.; Shaw, D. E. D. E. Shaw Research, Technical Report DESRES/TR--2014-01, 2014. (41) Liu, S.; Wu, Y.; Lin, T.; Abel, R.; Redmann, J.; Summa, C.; Jaber, V.; Lim, N.; Mobley, D. J. Comput.-Aided Mol. Des. 2013, 27, 755. (42) Prime, version 3.8; Schrodinger Inc., New York, 2014. (43) Glide, version 6.5; Schrodinger Inc., New York, 2014.

affinity. This is particularly significant when considering whether to make an otherwise highly attractive molecule that may be synthetically challenging. If such a molecule is predicted to achieve the project potency targets by reliable FEP calculations, this substantially reduces the risk of taking on such synthetic challenges. In addition, the elimination of compounds unlikely to meet project potency targets frees resources to focus on more promising compounds. Thus, extensive deployment of FEP in a drug discovery project not only will reduce the number of compounds that are made with inadequate activity but also may facilitate significant leaps in chemical space that otherwise would not have been taken, leading to more rapid completion of difficult projects, with potentially superior molecules as an end result.



ASSOCIATED CONTENT

S Supporting Information *

Additional text and equations with details of REST region selection algorithm, simulations, OPLS2.1 force field, conversion of calculated ΔΔG values to ΔG values, expected FEP prediction accuracy in prospective studies, expected correlation coefficient between FEP-predicted binding affinities and experimental values and between two independent experimental measurements, experimental binding affinity data, and input structures for FEP calculations; three tables listing representative p38 ligand pairs used in FEP/REST calculations, comparison of RMSE for FEP-predicted binding affinities for all 330 pairs of ligands and RMSE assuming all ligands are equally potent, and number of OPLS2.1 missing torsions identified for three ligand sets; and two figures showing error distribution of FEP-predicted relative binding free energies compared to experimental data and its fitting by a Gaussian function and histograms of the RMSE in force-field relative energies (pdf). Experimental and predicted ΔG values and errors for all eight data sets (xls). This material is available free of charge via the Internet at http://pubs.acs.org.



AUTHOR INFORMATION

Corresponding Author

*[email protected] Notes

The authors declare the following competing financial interest(s): D.L.M., W.L.J., and B.J.B. are consultants to Schrodinger, Inc. and are on its Scientific Advisory Board. R.A.F. has a significant financial stake in, is a consultant for, and is on the Scientific Advisory Board of Schrodinger, Inc.



REFERENCES

(1) Jorgensen, W. L. Acc. Chem. Res. 2009, 42, 724. (2) Jorgensen, W. L. Science 2004, 303, 1813. (3) Homeyer, N.; Stoll, F.; Hillisch, A.; Gohlke, H. J. Chem. Theory Comput. 2014, 10, 3331. (4) Free Energy Calculations: Theory and Applications in Chemistry and Biology; Chipot, C., Pohorille, A., Eds.; Springer Series in Chemical Physics, Vol. 86; Springer: Berlin and Heidelberg, Germany, 2007. (5) Chodera, J. D.; Mobley, D. L.; Shirts, M. R.; Dixon, R. W.; Branson, K.; Pande, V. S. Curr. Opin. Struct. Biol. 2011, 21, 150. (6) Knight, J. L.; Brooks, C. L. J. Comput. Chem. 2009, 30, 1692. (7) Zheng, L.; Chen, M.; Yang, W. Proc. Natl. Acad. Sci. U.S.A. 2008, 105, 20227. (8) Gallicchio, E.; Levy, R. M. Curr. Opin. Struct. Biol. 2011, 21, 161. (9) Hansen, N.; van Gunsteren, W. F. J. Chem. Theory Comput. 2014, 10, 2632. 2702

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703

Article

Journal of the American Chemical Society (44) Brown, S. P.; Muchmore, S. W.; Hajduk, P. J. Drug Discovery Today 2009, 14, 420. (45) Baum, B.; Mohamed, M.; Zayed, M.; Gerlach, C.; Heine, A.; Hangauer, D.; Klebe, G. J. Mol. Biol. 2009, 390, 56. (46) Cumming, J. N.; Smith, E. M.; Wang, L.; Misiaszek, J.; Durkin, J.; Pan, J.; Iserloh, U.; Wu, Y.; Zhu, Z.; Strickland, C.; Voigt, J.; Chen, X.; Kennedy, M. E.; Kuvelkar, R.; Hyde, L. A.; Cox, K.; Favreau, L.; Czarniecki, M. F.; Greenlee, W. J.; McKittrick, B. A.; Parker, E. M.; Stamford, A. W. Bioorg. Med. Chem. Lett. 2012, 22, 2444. (47) Hardcastle, I. R.; Arris, C. E.; Bentley, J.; Boyle, F. T.; Chen, Y.; Curtin, N. J.; Endicott, J. A.; Gibson, A. E.; Golding, B. T.; Griffin, R. J.; Jewsbury, P.; Menyerol, J.; Mesguiche, V.; Newell, D. R.; Noble, M. E.; Pratt, D. J.; Wang, L. Z.; Whitfield, H. J. J. Med. Chem. 2004, 47, 3710. (48) Szczepankiewicz, B. G.; Kosogof, C.; Nelson, L. T. J.; Liu, G.; Liu, B.; Zhao, H.; Serby, M. D.; Xin, Z.; Liu, M.; Gum, R. J.; Haasch, D. L.; Wang, S.; Clampit, J. E.; Johnson, E. F.; Lubben, T. H.; Stashko, M. A.; Olejniczak, E. T.; Sun, C.; Dorwin, S. A.; Haskins, K.; AbadZapatero, C.; Fry, E. H.; Hutchins, C. W.; Sham, H. L.; Rondinone, C. M.; Trevillyan, J. M. J. Med. Chem. 2006, 49, 3563. (49) Friberg, A.; Vigil, D.; Zhao, B.; Daniels, R. N.; Burke, J. P.; Garcia-Barrantes, P. M.; Camper, D.; Chauder, B. A.; Lee, T.; Olejniczak, E. T.; Fesik, S. W. J. Med. Chem. 2012, 56, 15. (50) Goldstein, D. M.; Soth, M.; Gabriel, T.; Dewdney, N.; Kuglstatter, A.; Arzeno, H.; Chen, J.; Bingenheimer, W.; Dalrymple, S. A.; Dunn, J.; Farrell, R.; Frauchiger, S.; La Fargue, J.; Ghate, M.; Graves, B.; Hill, R. J.; Li, F.; Litman, R.; Loe, B.; McIntosh, J.; McWeeney, D.; Papp, E.; Park, J.; Reese, H. F.; Roberts, R. T.; Rotstein, D.; San Pablo, B.; Sarma, K.; Stahl, M.; Sung, M.-L.; Suttman, R. T.; Sjogren, E. B.; Tan, Y.; Trejo, A.; Welch, M.; Weller, P.; Wong, B. R.; Zecic, H. J. Med. Chem. 2011, 54, 2255. (51) Wilson, D. P.; Wan, Z.-K.; Xu, W.-X.; Kirincich, S. J.; Follows, B. C.; Joseph-Mccarthy, D.; Foreman, K.; Moretto, A.; Wu, J.; Zhu, M.; Binnun, E.; Zhang, Y.-L.; Tam, M.; Erbe, D. V.; Tobin, J.; Xu, X.; Leung, L.; Shilling, A.; Tam, S. Y.; Mansour, T. S.; Lee, J. J. Med. Chem. 2007, 50, 4681. (52) Liang, J.; Tsui, V.; Van Abbema, A.; Bao, L.; Barrett, K.; Beresini, M.; Berezhkovskiy, L.; Blair, W. S.; Chang, C.; Driscoll, J.; Eigenbrot, C.; Ghilardi, N.; Gibbons, P.; Halladay, J.; Johnson, A.; Kohli, P. B.; Lai, Y.; Liimatta, M.; Mantik, P.; Menghrajani, K.; Murray, J.; Sambrone, A.; Xiao, Y.; Shia, S.; Shin, Y.; Smith, J.; Sohn, S.; Stanley, M.; Ultsch, M.; Zhang, B.; Wu, L. C.; Magnuson, S. Eur. J. Med. Chem. 2013, 67, 175. (53) Liang, J.; van Abbema, A.; Balazs, M.; Barrett, K.; Berezhkovsky, L.; Blair, W.; Chang, C.; Delarosa, D.; DeVoss, J.; Driscoll, J.; Eigenbrot, C.; Ghilardi, N.; Gibbons, P.; Halladay, J.; Johnson, A.; Kohli, P. B.; Lai, Y.; Liu, Y.; Lyssikatos, J.; Mantik, P.; Menghrajani, K.; Murray, J.; Peng, I.; Sambrone, A.; Shia, S.; Shin, Y.; Smith, J.; Sohn, S.; Tsui, V.; Ultsch, M.; Wu, L. C.; Xiao, Y.; Yang, W.; Young, J.; Zhang, B.; Zhu, B.-y.; Magnuson, S. J. Med. Chem. 2013, 56, 4521.

2703

DOI: 10.1021/ja512751q J. Am. Chem. Soc. 2015, 137, 2695−2703